Incorporating Donor Lung CT Images into Machine Learning Models to Predict Severe Primary Graft Dysfunction after Lung Transplantation - PROJECT SUMMARY/ABSTRACT Primary graft dysfunction (PGD) is a form of acute lung injury that occurs in the first 72 hours after lung transplantation. PGD is the leading cause of morbidity and mortality in the first 30 days and is strongly associated with an increased risk of chronic rejection and death beyond the first year after transplantation. Donor, recipient, and surgical risk factors for PGD have been identified, and clinicians make decisions about donor offer acceptance based on these risk factors to avoid severe PGD. However, how different donor and recipient risk factors interact to impact the risk of PGD is unclear. We previously built a machine learning (ML) algorithm to predict the risk of severe PGD using clinical data available at the time of donor offer. However, this model was based on single-center data, and its performance could be improved by capturing additional key parameters using a multi-center cohort. Furthermore, the model did not include donor lung CT images through imaging is a cornerstone of clinical decision-making. We hypothesize that donor lung CT images encode valuable information that could be used to estimate the risk of severe PGD and that better risk estimates could be achieved by fusing donor and recipient clinical data with donor CT imaging data. To test this hypothesis, we will perform an observational study at five lung transplant centers in the US. We will use a retrospective cohort of bilateral lung transplant recipients from the sites to build ML models to predict severe PGD, and concurrently, enroll lung transplant recipients in a prospective cohort, which will serve for model validation. To determine the characteristics of donor CT images associated with severe PGD, we will use the attention-based multiple instance learning (MIL) method. We will divide CT scans into 3D patches of 32×32×32 voxels per patch, and severe PGD will serve as the pseudo-label for all 3D patches. Both the original CT scans and the augmented sub-volumes will be processed through separate Swin Transformer encoders and combined into three “heads” used to compute the inpaint loss, contrastive loss, and bounding box detection loss. Gradient-weighted class activation mapping (Grad-CAM) will be applied to the attention scores of the 3D patches to visualize prioritized lung regions. Attention masks generated by the MIL model will be examined by a thoracic radiology expert to identify and list imaging biomarkers predictive of severe PGD. We will then integrate the clinical and imaging data using a multimodal learning strategy. The clinical and MIL-identified featuresets will be aligned using separate Multi-Layer Perceptron (MLP) layers before concatenation and passed through another series of MLP layers. To avoid overfitting, clinical and MIL-identified features will be combined in an element-wise fashion with their independent feature vectors before being passed through the final MLP layers. This project will develop a multi-dimensional clinical decision-making tool to predict the risk of severe PGD. This can enhance donor- recipient matching and clinical preparedness for immediate post-transplant care to improve early outcomes after lung transplantation.
